Since previous work showed that it is difficult to forecast dry spells in Malawi, this document explores the option of using monthly precipitation as a proxy for the occurrence of dry spells and the forecasting skill for monthly precipitation. The precipitation patterns differ per region in Malawi. Since 41 out of 44 month-admin2 combinations with a dry spell occurred in the Southern region, this analysis solely focuses on this region to prevent weakening of signals due to the differences in climates between the regions.
This analysis
The figure below shows the monthly precipitation, colored by whether a dry spell started during that month. From this figure we can see that
We can see that precipitation patterns clearly differ by month, which is expected. Moreover, the figures show that the separability of the occurrence of a dry spell differs by month. This becomes more clear when looking at a boxplot
In order to use the monthly precipitation as a proxy for the occurence of a dry spell, we need to set a threshold on the amount of monthly precipitation. I.e. if the precipitation is forecasted to be lower than X, we trigger and else we don’t.
One method to get a sense for this, is by investigating a range of thresholds, and for each of them looking at the percentage of months with a dry spell that are below the threshold, and the percentage of months without a dry spell that are above the threshold. This is shown in the figure below. The figure shows that the trade-off between sensitivity and specificity depends on the month. Moreover, there is better separability during January and February than during December.
Since a trigger becomes too complex when setting a threshold for each month separately, we decide to set one threshold for all months. This is chosen by choosing the intersection of the two lines computed over all months, which is at 180mm. The graph below illustrates this.
To further understand the co-occurrence of low amounts of monthly precipitation and dry spells, it is important to look at the temporal distribution. To do so, the below figure shows a heatmap indicating the months with less than 180 mm of precipitation and the occurrence of dry spells. From this figure we can conclude that most dry spells overlap with most of <=180 mm, except a few occurrences in December and one in February 2002. It also becomes clear that <=180 mm monthly precipitation occurs in years that no dry spell was identified. Thus, with this threshold one would trigger more often than the occurrences of dry spells.
To conclude, the confusion matrix with a threshold of 180 mm is as follows. Thus 78% of the dry spells would be identified, and in 33% of the months with less than 180 mm of precipitation, a dry spell occurred.
Within this pilot, the accepted definition of a dry spell is that of <=2mm cumulative precipitation during 14 consecutive days. However, observational and forecast products often have trouble detecting such small amounts of rainfall and therefore many practitioners before us have defined dry spells based on dry days. Meaning that a dry spell consists of at least 14 consecutive dry days. The definition of a dry day differs between projects, but we take a commonly used definition of 4mm/day.
We compute the same analysis with this different definition, and find that the optimal threshold is also at 180mm. As shown below.
From the heatmap we can conclude that there are slightly more occurrences of dry spells than when allowing max 2mm of cumulative precipitation, which is as expected.
NOTE: the missing of dry spells during December compared to the other definition, is due to the fact that with this definition only dry spells during the rainy season were included –> to be improved.
When looking at the confusion matrix, we can see slightly more agreement. 90% of the dry spells would be identified, and in 43% of the months with less than 180 mm of precipitation, a dry spell occurred.
NOTE: these numbers are missing the dry spells outside the rainy season, which will probably lower the 90%!
To support our discussion on severity, here is the heatmap on admin2 with <=2mm cumulative rainfall